Extracting Terrain Features from Range Images for Autonomous Random Stepfield Traversal Raymond Sheh, M. Waleed Kadous, Claude Sammut ARC Centre of Excellence for Autonomous Systems School of Computer Science and Engineering The University of New South Wales Sydney NSW 2052 Australia [rsheh|waleed|claude]@cse.unsw.edu.au Bernhard Hengst NICTA School of Computer Science and Engineering The University of New South Wales Sydney NSW 2052 Australia bernhard.hengst@nicta.com.au Abstract — One of the challenges of rescue robotics is to create robots that can autonomously traverse rough, unstructured terrain. Although mechanical engineering can produce very capable robots, mechanical engineering alone will not drive them. In this paper, we present a terrain feature extractor that can be taught to find significant features in range images of terrain around a robot from a human expert. This novel approach has the advantage that it potentially allows the human expert’s knowledge to be captured rapidly. A terrain model is generated from the many points in the range sensor data. Techniques from the field of knowledge acquisition are then used to find patterns in the terrain model. A knowledge acquisition system can then be taught to drive a robot in unstructured terrain based on these features. We evaluate the performance of the initial stages of the feature extractor on a real robot, traversing NIST specification red stepfields. I. I NTRODUCTION One of the challenges of rescue robotics is to create robots that can autonomously traverse rubble and rough terrain at a disaster site. Mechanical engineering can help in this regard and competitions like RoboCup Rescue see many very capable designs for the purpose of traversing rough terrain on a scale seen in search and rescue environments. However, mechanics alone won’t drive a robot. Much of the terrain that rescue robots can be expected to operate in consists of highly unstructured features that cannot be easily classified into traversable or non-traversable, even if they could be identified. At a human scale, this would be like scrambling over a rocky seawall where the traversability of the various rocks is neither obvious nor constant. This traversability would also depend more heavily on the ability of the human (height, flexibility, strength, etc.) than in the case of walking along, say, a forest path. Our aim is to create robots that are able to traverse these highly unstructured terrains. We use NIST Random Stepfields [2] as an analog for real rubble. Stepfields are designed to be easily reproduced, and yet behave in a similar way to real rubble. To standardise difficulty, NIST publishes a set of randomly generated patterns in various sizes. Our experiments are run on NIST Red stepfields which consist of blocks around 90mm per side and up to 400mm high. This paper describes the development of a set of features that can compactly represent the terrain around the robot, as sensed by the robot’s range imager. These features are used by the decision making processes on the robot in order to decide on the appropriate action to take in order to traverse the terrain. A. Platform and Sensors Our robot, CASTER3, is based on a Yujin Robotics Robhaz DT-3 EOD robot, and was our entry in the 2005 and 2006 RoboCup Rescue Robot League competitions where we came 3rd and reached the semi-finals respectively. CASTER3 was designed for stairclimbing and overcoming obstacles in an urban area, however its two pairs of tracks and articulated body give it reasonable performance on NIST Red stepfields. Our selection of robot and terrain is such that one cannot drive the robot over the stepfield as if it were flat ground. Doing so would result in the robot being trapped. However, when driven with care, it is often possible to traverse the terrain. Thus it is a good test for autonomous agent systems that are to sense the terrain and drive carefully. Our primary sensor for terrain characterisation is the CSEM SwissRanger SR-3100 range imager [4]. This camera-like device provides a 176x144 array of pixels, each pixel being a distance value with a field of view of approximately 40 ◦ . Compared to stereo vision, this 3D data is low in distortion, dense and invariant to lighting and surface texture. Unlike scanning lasers, the whole field of view is sensed in one go. An XSens MTi heading-attitude sensor allows us to rotate the 3D data to the horizontal and provides an estimate for yaw. B. Related Work Several groups have tackled the issue of 3D terrain represen- tation, mostly for autonomous road vehicles. In our applica- tion, the size and complexity of obstacles can be considerably greater. It is not unusual for our robot to need to drive over multiple obstacles that are comparable to the size of the robot itself. Our focus is on how to drive over these obstacles rather than simply classifying and avoiding them. Representations such as those used in [6] process point- clouds sensed at a particular instant and do not make use of an ongoing map. Obstacles are segmented based on their deviation from the driving surface, which need not be level or flat. The thresholding of these deviations is governed by